Analysis and Detection of Orange Images Based on Improved Faster R-CNN Algorithm and Feature Data Analysis

摘要

Orange is one of the fruits with the largest planting area in China, and the harvesting work is huge in the harvest season. Therefore, the orange picking robot is often used to replace workers for picking work. In order to enable the picking robot to better detect orange fruits in the natural environment, we established an improved faster R-CNN detection model based on the growth characteristics of oranges. First, to reduce the orange detection model, VGG16 was replaced by MobileNetv2 as the feature extraction network of this algorithm. Secondly, to make the model fit better, the size of the anchor frame was fine-tuned. Finally, aiming at the problem that trees of orange are easy to overlap and block each other, a double-NMS algorithm was designed to replace NMS and remove redundant rectangular frames. The experimental results show that the improved algorithm has higher detection accuracy and lower missed detection rate, which provides a new method for improving the intelligence of the picking robot.